Support Chat Completions wire format for custom OpenAI-compatible providers

[diogenesc]

Problem

The codex runtime exclusively uses the OpenAI Responses API (/v1/responses) wire format. Custom OpenAI-compatible model serving platforms — such as Databricks Model Serving, vLLM, LiteLLM, and others — typically only implement the Chat Completions API (/v1/chat/completions).

This makes the new inline engine definition with custom providers (from #20416 / #20473) non-functional for these platforms: the OAuth token exchange and provider configuration work correctly at the compiler level, but the codex runtime sends requests in a format the endpoint cannot serve.

Evidence from a real workflow run

Using the inline engine definition with oauth-client-credentials against a custom OpenAI-compatible endpoint:

1. OAuth token exchange succeeds

The compiler correctly obtains a token and populates OPENAI_API_KEY:

✅ OPENAI_API_KEY: Configured (using as fallback for CODEX_API_KEY)

2. Model name is ignored — codex resolves to its default

Despite GH_AW_MODEL_AGENT_CODEX being set to the custom model name, codex CLI resolves to its hardcoded default:

GH_AW_MODEL_AGENT_CODEX: <custom-model-name>

But the session init log shows:

DEBUG session_init: codex_core::codex: Configuring session:
  model=gpt-5.3-codex;
  provider=ModelProviderInfo {
    name: "OpenAI",
    base_url: Some("http://172.30.0.30:10000/v1"),
    wire_api: Responses,
    ...
  }

The custom model name was replaced with gpt-5.3-codex, and wire_api is hardcoded to Responses.

3. Codex sends POST to /v1/responses — endpoint rejects it

TRACE: codex_client::transport: POST to http://172.30.0.30:10000/v1/responses
  {"model":"gpt-5.3-codex", ...}

The API proxy forwards this to the custom endpoint, which only serves /v1/chat/completions. The request fails silently:

[WARN] Command completed with exit code: 1
Process exiting with code: 1

4. Agent produces no output

OUTPUT_TYPES: (empty)
HAS_PATCH: false
Detection skipped: no agent outputs or patches to analyze

Root cause

There are two independent issues:

1. Wire format mismatch: The codex runtime hardcodes wire_api: Responses. Custom providers using Chat Completions cannot serve /v1/responses requests. The request.path-template field from Phase 4: Add AuthDefinition and RequestShape for provider-owned auth and request shaping #20473 only rewrites the URL path — it does not translate the request/response body format.

2. Model name override ignored: The provider.model field from the frontmatter is passed as GH_AW_MODEL_AGENT_CODEX but codex CLI does not honor it, falling back to gpt-5.3-codex. This means even if the wire format were compatible, the wrong model would be requested.

Proposal

Add a request.wire-format field (or equivalent) to the inline engine definition that controls which wire protocol the runtime uses:

engine:
  runtime:
    id: codex
  provider:
    id: my-provider
    model: my-custom-model
    request:
      wire-format: chat-completions  # "responses" (default) | "chat-completions"
    auth:
      strategy: oauth-client-credentials
      token-url: https://example.com/oidc/v1/token
      client-id: MY_CLIENT_ID
      client-secret: MY_CLIENT_SECRET

When wire-format: chat-completions is set, the runtime should:

1. Send requests to /v1/chat/completions instead of /v1/responses
2. Translate the request body to Chat Completions format
3. Parse the Chat Completions response back into the internal format
4. Honor the provider.model field in API requests

Alternatively, the API proxy layer (172.30.0.30:10000) could perform this translation transparently when the engine definition declares chat-completions wire format.

Impact

This blocks adoption of GitHub Agentic Workflows for any organization that mandates use of self-hosted or third-party OpenAI-compatible inference platforms. The engine enhancement architecture from #20416 already supports custom providers and auth — wire format is the last missing piece.

Versions

• gh-aw: v0.58.0
• codex CLI: v0.114.0
• API proxy: gh-aw-firewall v0.23.0

Related

• GitHub Agentic Workflow Engine Enhancement Proposal #20416 — Engine Enhancement Proposal (introduced provider/runtime separation)
• Phase 4: Add AuthDefinition and RequestShape for provider-owned auth and request shaping #20473 — Phase 4: Auth & Request Shaping (merged 2026-03-11)